spot nuclei. speed cures - github pages · spot nuclei. speed cures chih-hui ho, chun-han yao,...

Spot Nuclei. Speed CuresChih-Hui Ho, Chun-Han Yao, Po-Ya Hsu,

Yao-Yuan Yang, Hsin-Yang Chen, Ying-Chuan Liao

1

Outline

● Introduction● Dataset● Approaches

○ Pre-processing○ CNN○ Post-processing

● Results● Future work

2

Introduction● Identifying the nuclei of cells is the starting point for most medical analyses ● It allows researchers to locate each individual cell and understand the

underlying biological processes● It costs considerable time and effort for human researchers to label the cells● Our goal is to advance medical discovery by automating the identification of

nuclei in diverse images

3

Dataset● Kaggle 2018 Data Science Bowl● 670 training images with 29462 segmentation masks● 65 testing images with no ground truth● Each image has different size

Example training images Example testing images 4

Framework overview

5

Approaches -- Pre-processing ● Several types of input images, such as RGB, grayscale and HSV are

experimented

● Pre-classify the source of the image into sub-classes

● Histogram equalization

● Different input sizes (128 - 512)

6

Approaches -- Physical Method● Treat the whole image itself as a distribution of light. Employ the intensity and

gradient as features to capture cells.

Original Image in Grayscale (Left) ; Features to Capture Cells (Right)

7

Approaches -- Computer Vision Method

Qualitative result of superpixel segmentation with SVM classification and GMM clustering. The input image, superpixel segmentation at multiple scale, and the predicted mask are shown from left to right.

8

Approaches -- CNN● UNet is designed for biomedical image segmentation, where the

precision of mask boundary is critical● Prediction error at the boundaries are heavily penalized● We found that changing the input size and post processing works better

UNet architectureWeighted penaltyGround truth 9

Approaches -- CNN● Modified UNet architecture● Instead of concatenating the upsampled feature map with original feature

map, an extra ‘high way’ path is provided● The pooling layer is also replaced by convolution layer with larger stride● This improves the accuracy by 6%

Convolution layers

10

Approaches -- CNN (DeepMask/SharpMask)● DeepMask

○ Introduced by FAIR (Facebook AI Research)○ Two outputs (Score, Mask)

11

Approaches -- CNN (DeepMask/SharpMask)● SharpMask● Inherited from DeepMask● Use Resnet instead of VGG● Additional refinement module● Experiment different head options

12

Approaches -- CNN (DeepMask/SharpMask)● Inference stage● Sliding window● Masks with high scores → Heat map

Test image Heat map Final mask

13

Approaches -- Post-processing

● Fill the holes in each connected components

● Gaussian Mixture Model (GMM) clustering○ The number of clusters is determined by log likelihood thresholding and the “elbow method”

● Remove the instances with irregular size and shape● Ensemble method

○ Collect result from several models○ Every pixel is voted and the majority rule is applied

Negative log likelihood

Number of clusters

14

Current Results● Top 12% in current Kaggle leaderboard

15

Future work● Implement architecture that includes both detection and segmentation, such

as mask RCNN● Figure out how a cell nucleus is defined, and then modify the formulation(s) to

improve the models

16